Commit | Line | Data |
---|---|---|
b20a3503 CL |
1 | /* |
2 | * Memory Migration functionality - linux/mm/migration.c | |
3 | * | |
4 | * Copyright (C) 2006 Silicon Graphics, Inc., Christoph Lameter | |
5 | * | |
6 | * Page migration was first developed in the context of the memory hotplug | |
7 | * project. The main authors of the migration code are: | |
8 | * | |
9 | * IWAMOTO Toshihiro <iwamoto@valinux.co.jp> | |
10 | * Hirokazu Takahashi <taka@valinux.co.jp> | |
11 | * Dave Hansen <haveblue@us.ibm.com> | |
cde53535 | 12 | * Christoph Lameter |
b20a3503 CL |
13 | */ |
14 | ||
15 | #include <linux/migrate.h> | |
b95f1b31 | 16 | #include <linux/export.h> |
b20a3503 | 17 | #include <linux/swap.h> |
0697212a | 18 | #include <linux/swapops.h> |
b20a3503 | 19 | #include <linux/pagemap.h> |
e23ca00b | 20 | #include <linux/buffer_head.h> |
b20a3503 | 21 | #include <linux/mm_inline.h> |
b488893a | 22 | #include <linux/nsproxy.h> |
b20a3503 | 23 | #include <linux/pagevec.h> |
e9995ef9 | 24 | #include <linux/ksm.h> |
b20a3503 CL |
25 | #include <linux/rmap.h> |
26 | #include <linux/topology.h> | |
27 | #include <linux/cpu.h> | |
28 | #include <linux/cpuset.h> | |
04e62a29 | 29 | #include <linux/writeback.h> |
742755a1 CL |
30 | #include <linux/mempolicy.h> |
31 | #include <linux/vmalloc.h> | |
86c3a764 | 32 | #include <linux/security.h> |
8a9f3ccd | 33 | #include <linux/memcontrol.h> |
4f5ca265 | 34 | #include <linux/syscalls.h> |
290408d4 | 35 | #include <linux/hugetlb.h> |
8e6ac7fa | 36 | #include <linux/hugetlb_cgroup.h> |
5a0e3ad6 | 37 | #include <linux/gfp.h> |
bf6bddf1 | 38 | #include <linux/balloon_compaction.h> |
b20a3503 | 39 | |
0d1836c3 MN |
40 | #include <asm/tlbflush.h> |
41 | ||
7b2a2d4a MG |
42 | #define CREATE_TRACE_POINTS |
43 | #include <trace/events/migrate.h> | |
44 | ||
b20a3503 CL |
45 | #include "internal.h" |
46 | ||
b20a3503 | 47 | /* |
742755a1 | 48 | * migrate_prep() needs to be called before we start compiling a list of pages |
748446bb MG |
49 | * to be migrated using isolate_lru_page(). If scheduling work on other CPUs is |
50 | * undesirable, use migrate_prep_local() | |
b20a3503 CL |
51 | */ |
52 | int migrate_prep(void) | |
53 | { | |
b20a3503 CL |
54 | /* |
55 | * Clear the LRU lists so pages can be isolated. | |
56 | * Note that pages may be moved off the LRU after we have | |
57 | * drained them. Those pages will fail to migrate like other | |
58 | * pages that may be busy. | |
59 | */ | |
60 | lru_add_drain_all(); | |
61 | ||
62 | return 0; | |
63 | } | |
64 | ||
748446bb MG |
65 | /* Do the necessary work of migrate_prep but not if it involves other CPUs */ |
66 | int migrate_prep_local(void) | |
67 | { | |
68 | lru_add_drain(); | |
69 | ||
70 | return 0; | |
71 | } | |
72 | ||
b20a3503 | 73 | /* |
894bc310 LS |
74 | * Add isolated pages on the list back to the LRU under page lock |
75 | * to avoid leaking evictable pages back onto unevictable list. | |
b20a3503 | 76 | */ |
e13861d8 | 77 | void putback_lru_pages(struct list_head *l) |
b20a3503 CL |
78 | { |
79 | struct page *page; | |
80 | struct page *page2; | |
b20a3503 | 81 | |
5733c7d1 RA |
82 | list_for_each_entry_safe(page, page2, l, lru) { |
83 | list_del(&page->lru); | |
84 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
85 | page_is_file_cache(page)); | |
86 | putback_lru_page(page); | |
87 | } | |
88 | } | |
89 | ||
90 | /* | |
91 | * Put previously isolated pages back onto the appropriate lists | |
92 | * from where they were once taken off for compaction/migration. | |
93 | * | |
94 | * This function shall be used instead of putback_lru_pages(), | |
95 | * whenever the isolated pageset has been built by isolate_migratepages_range() | |
96 | */ | |
97 | void putback_movable_pages(struct list_head *l) | |
98 | { | |
99 | struct page *page; | |
100 | struct page *page2; | |
101 | ||
b20a3503 | 102 | list_for_each_entry_safe(page, page2, l, lru) { |
31caf665 NH |
103 | if (unlikely(PageHuge(page))) { |
104 | putback_active_hugepage(page); | |
105 | continue; | |
106 | } | |
e24f0b8f | 107 | list_del(&page->lru); |
a731286d | 108 | dec_zone_page_state(page, NR_ISOLATED_ANON + |
6c0b1351 | 109 | page_is_file_cache(page)); |
117aad1e | 110 | if (unlikely(isolated_balloon_page(page))) |
bf6bddf1 RA |
111 | balloon_page_putback(page); |
112 | else | |
113 | putback_lru_page(page); | |
b20a3503 | 114 | } |
b20a3503 CL |
115 | } |
116 | ||
0697212a CL |
117 | /* |
118 | * Restore a potential migration pte to a working pte entry | |
119 | */ | |
e9995ef9 HD |
120 | static int remove_migration_pte(struct page *new, struct vm_area_struct *vma, |
121 | unsigned long addr, void *old) | |
0697212a CL |
122 | { |
123 | struct mm_struct *mm = vma->vm_mm; | |
124 | swp_entry_t entry; | |
0697212a CL |
125 | pmd_t *pmd; |
126 | pte_t *ptep, pte; | |
127 | spinlock_t *ptl; | |
128 | ||
290408d4 NH |
129 | if (unlikely(PageHuge(new))) { |
130 | ptep = huge_pte_offset(mm, addr); | |
131 | if (!ptep) | |
132 | goto out; | |
133 | ptl = &mm->page_table_lock; | |
134 | } else { | |
6219049a BL |
135 | pmd = mm_find_pmd(mm, addr); |
136 | if (!pmd) | |
290408d4 | 137 | goto out; |
500d65d4 AA |
138 | if (pmd_trans_huge(*pmd)) |
139 | goto out; | |
0697212a | 140 | |
290408d4 | 141 | ptep = pte_offset_map(pmd, addr); |
0697212a | 142 | |
486cf46f HD |
143 | /* |
144 | * Peek to check is_swap_pte() before taking ptlock? No, we | |
145 | * can race mremap's move_ptes(), which skips anon_vma lock. | |
146 | */ | |
290408d4 NH |
147 | |
148 | ptl = pte_lockptr(mm, pmd); | |
149 | } | |
0697212a | 150 | |
0697212a CL |
151 | spin_lock(ptl); |
152 | pte = *ptep; | |
153 | if (!is_swap_pte(pte)) | |
e9995ef9 | 154 | goto unlock; |
0697212a CL |
155 | |
156 | entry = pte_to_swp_entry(pte); | |
157 | ||
e9995ef9 HD |
158 | if (!is_migration_entry(entry) || |
159 | migration_entry_to_page(entry) != old) | |
160 | goto unlock; | |
0697212a | 161 | |
0697212a CL |
162 | get_page(new); |
163 | pte = pte_mkold(mk_pte(new, vma->vm_page_prot)); | |
c3d16e16 CG |
164 | if (pte_swp_soft_dirty(*ptep)) |
165 | pte = pte_mksoft_dirty(pte); | |
0697212a CL |
166 | if (is_write_migration_entry(entry)) |
167 | pte = pte_mkwrite(pte); | |
3ef8fd7f | 168 | #ifdef CONFIG_HUGETLB_PAGE |
be7517d6 | 169 | if (PageHuge(new)) { |
290408d4 | 170 | pte = pte_mkhuge(pte); |
be7517d6 TL |
171 | pte = arch_make_huge_pte(pte, vma, new, 0); |
172 | } | |
3ef8fd7f | 173 | #endif |
c2cc499c | 174 | flush_dcache_page(new); |
0697212a | 175 | set_pte_at(mm, addr, ptep, pte); |
04e62a29 | 176 | |
290408d4 NH |
177 | if (PageHuge(new)) { |
178 | if (PageAnon(new)) | |
179 | hugepage_add_anon_rmap(new, vma, addr); | |
180 | else | |
181 | page_dup_rmap(new); | |
182 | } else if (PageAnon(new)) | |
04e62a29 CL |
183 | page_add_anon_rmap(new, vma, addr); |
184 | else | |
185 | page_add_file_rmap(new); | |
186 | ||
187 | /* No need to invalidate - it was non-present before */ | |
4b3073e1 | 188 | update_mmu_cache(vma, addr, ptep); |
e9995ef9 | 189 | unlock: |
0697212a | 190 | pte_unmap_unlock(ptep, ptl); |
e9995ef9 HD |
191 | out: |
192 | return SWAP_AGAIN; | |
0697212a CL |
193 | } |
194 | ||
04e62a29 CL |
195 | /* |
196 | * Get rid of all migration entries and replace them by | |
197 | * references to the indicated page. | |
198 | */ | |
199 | static void remove_migration_ptes(struct page *old, struct page *new) | |
200 | { | |
e9995ef9 | 201 | rmap_walk(new, remove_migration_pte, old); |
04e62a29 CL |
202 | } |
203 | ||
0697212a CL |
204 | /* |
205 | * Something used the pte of a page under migration. We need to | |
206 | * get to the page and wait until migration is finished. | |
207 | * When we return from this function the fault will be retried. | |
0697212a | 208 | */ |
30dad309 NH |
209 | static void __migration_entry_wait(struct mm_struct *mm, pte_t *ptep, |
210 | spinlock_t *ptl) | |
0697212a | 211 | { |
30dad309 | 212 | pte_t pte; |
0697212a CL |
213 | swp_entry_t entry; |
214 | struct page *page; | |
215 | ||
30dad309 | 216 | spin_lock(ptl); |
0697212a CL |
217 | pte = *ptep; |
218 | if (!is_swap_pte(pte)) | |
219 | goto out; | |
220 | ||
221 | entry = pte_to_swp_entry(pte); | |
222 | if (!is_migration_entry(entry)) | |
223 | goto out; | |
224 | ||
225 | page = migration_entry_to_page(entry); | |
226 | ||
e286781d NP |
227 | /* |
228 | * Once radix-tree replacement of page migration started, page_count | |
229 | * *must* be zero. And, we don't want to call wait_on_page_locked() | |
230 | * against a page without get_page(). | |
231 | * So, we use get_page_unless_zero(), here. Even failed, page fault | |
232 | * will occur again. | |
233 | */ | |
234 | if (!get_page_unless_zero(page)) | |
235 | goto out; | |
0697212a CL |
236 | pte_unmap_unlock(ptep, ptl); |
237 | wait_on_page_locked(page); | |
238 | put_page(page); | |
239 | return; | |
240 | out: | |
241 | pte_unmap_unlock(ptep, ptl); | |
242 | } | |
243 | ||
30dad309 NH |
244 | void migration_entry_wait(struct mm_struct *mm, pmd_t *pmd, |
245 | unsigned long address) | |
246 | { | |
247 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
248 | pte_t *ptep = pte_offset_map(pmd, address); | |
249 | __migration_entry_wait(mm, ptep, ptl); | |
250 | } | |
251 | ||
252 | void migration_entry_wait_huge(struct mm_struct *mm, pte_t *pte) | |
253 | { | |
254 | spinlock_t *ptl = &(mm)->page_table_lock; | |
255 | __migration_entry_wait(mm, pte, ptl); | |
256 | } | |
257 | ||
b969c4ab MG |
258 | #ifdef CONFIG_BLOCK |
259 | /* Returns true if all buffers are successfully locked */ | |
a6bc32b8 MG |
260 | static bool buffer_migrate_lock_buffers(struct buffer_head *head, |
261 | enum migrate_mode mode) | |
b969c4ab MG |
262 | { |
263 | struct buffer_head *bh = head; | |
264 | ||
265 | /* Simple case, sync compaction */ | |
a6bc32b8 | 266 | if (mode != MIGRATE_ASYNC) { |
b969c4ab MG |
267 | do { |
268 | get_bh(bh); | |
269 | lock_buffer(bh); | |
270 | bh = bh->b_this_page; | |
271 | ||
272 | } while (bh != head); | |
273 | ||
274 | return true; | |
275 | } | |
276 | ||
277 | /* async case, we cannot block on lock_buffer so use trylock_buffer */ | |
278 | do { | |
279 | get_bh(bh); | |
280 | if (!trylock_buffer(bh)) { | |
281 | /* | |
282 | * We failed to lock the buffer and cannot stall in | |
283 | * async migration. Release the taken locks | |
284 | */ | |
285 | struct buffer_head *failed_bh = bh; | |
286 | put_bh(failed_bh); | |
287 | bh = head; | |
288 | while (bh != failed_bh) { | |
289 | unlock_buffer(bh); | |
290 | put_bh(bh); | |
291 | bh = bh->b_this_page; | |
292 | } | |
293 | return false; | |
294 | } | |
295 | ||
296 | bh = bh->b_this_page; | |
297 | } while (bh != head); | |
298 | return true; | |
299 | } | |
300 | #else | |
301 | static inline bool buffer_migrate_lock_buffers(struct buffer_head *head, | |
a6bc32b8 | 302 | enum migrate_mode mode) |
b969c4ab MG |
303 | { |
304 | return true; | |
305 | } | |
306 | #endif /* CONFIG_BLOCK */ | |
307 | ||
b20a3503 | 308 | /* |
c3fcf8a5 | 309 | * Replace the page in the mapping. |
5b5c7120 CL |
310 | * |
311 | * The number of remaining references must be: | |
312 | * 1 for anonymous pages without a mapping | |
313 | * 2 for pages with a mapping | |
266cf658 | 314 | * 3 for pages with a mapping and PagePrivate/PagePrivate2 set. |
b20a3503 | 315 | */ |
36bc08cc | 316 | int migrate_page_move_mapping(struct address_space *mapping, |
b969c4ab | 317 | struct page *newpage, struct page *page, |
a6bc32b8 | 318 | struct buffer_head *head, enum migrate_mode mode) |
b20a3503 | 319 | { |
7039e1db | 320 | int expected_count = 0; |
7cf9c2c7 | 321 | void **pslot; |
b20a3503 | 322 | |
6c5240ae | 323 | if (!mapping) { |
0e8c7d0f | 324 | /* Anonymous page without mapping */ |
6c5240ae CL |
325 | if (page_count(page) != 1) |
326 | return -EAGAIN; | |
78bd5209 | 327 | return MIGRATEPAGE_SUCCESS; |
6c5240ae CL |
328 | } |
329 | ||
19fd6231 | 330 | spin_lock_irq(&mapping->tree_lock); |
b20a3503 | 331 | |
7cf9c2c7 NP |
332 | pslot = radix_tree_lookup_slot(&mapping->page_tree, |
333 | page_index(page)); | |
b20a3503 | 334 | |
edcf4748 | 335 | expected_count = 2 + page_has_private(page); |
e286781d | 336 | if (page_count(page) != expected_count || |
29c1f677 | 337 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { |
19fd6231 | 338 | spin_unlock_irq(&mapping->tree_lock); |
e23ca00b | 339 | return -EAGAIN; |
b20a3503 CL |
340 | } |
341 | ||
e286781d | 342 | if (!page_freeze_refs(page, expected_count)) { |
19fd6231 | 343 | spin_unlock_irq(&mapping->tree_lock); |
e286781d NP |
344 | return -EAGAIN; |
345 | } | |
346 | ||
b969c4ab MG |
347 | /* |
348 | * In the async migration case of moving a page with buffers, lock the | |
349 | * buffers using trylock before the mapping is moved. If the mapping | |
350 | * was moved, we later failed to lock the buffers and could not move | |
351 | * the mapping back due to an elevated page count, we would have to | |
352 | * block waiting on other references to be dropped. | |
353 | */ | |
a6bc32b8 MG |
354 | if (mode == MIGRATE_ASYNC && head && |
355 | !buffer_migrate_lock_buffers(head, mode)) { | |
b969c4ab MG |
356 | page_unfreeze_refs(page, expected_count); |
357 | spin_unlock_irq(&mapping->tree_lock); | |
358 | return -EAGAIN; | |
359 | } | |
360 | ||
b20a3503 CL |
361 | /* |
362 | * Now we know that no one else is looking at the page. | |
b20a3503 | 363 | */ |
7cf9c2c7 | 364 | get_page(newpage); /* add cache reference */ |
b20a3503 CL |
365 | if (PageSwapCache(page)) { |
366 | SetPageSwapCache(newpage); | |
367 | set_page_private(newpage, page_private(page)); | |
368 | } | |
369 | ||
7cf9c2c7 NP |
370 | radix_tree_replace_slot(pslot, newpage); |
371 | ||
372 | /* | |
937a94c9 JG |
373 | * Drop cache reference from old page by unfreezing |
374 | * to one less reference. | |
7cf9c2c7 NP |
375 | * We know this isn't the last reference. |
376 | */ | |
937a94c9 | 377 | page_unfreeze_refs(page, expected_count - 1); |
7cf9c2c7 | 378 | |
0e8c7d0f CL |
379 | /* |
380 | * If moved to a different zone then also account | |
381 | * the page for that zone. Other VM counters will be | |
382 | * taken care of when we establish references to the | |
383 | * new page and drop references to the old page. | |
384 | * | |
385 | * Note that anonymous pages are accounted for | |
386 | * via NR_FILE_PAGES and NR_ANON_PAGES if they | |
387 | * are mapped to swap space. | |
388 | */ | |
389 | __dec_zone_page_state(page, NR_FILE_PAGES); | |
390 | __inc_zone_page_state(newpage, NR_FILE_PAGES); | |
99a15e21 | 391 | if (!PageSwapCache(page) && PageSwapBacked(page)) { |
4b02108a KM |
392 | __dec_zone_page_state(page, NR_SHMEM); |
393 | __inc_zone_page_state(newpage, NR_SHMEM); | |
394 | } | |
19fd6231 | 395 | spin_unlock_irq(&mapping->tree_lock); |
b20a3503 | 396 | |
78bd5209 | 397 | return MIGRATEPAGE_SUCCESS; |
b20a3503 | 398 | } |
b20a3503 | 399 | |
290408d4 NH |
400 | /* |
401 | * The expected number of remaining references is the same as that | |
402 | * of migrate_page_move_mapping(). | |
403 | */ | |
404 | int migrate_huge_page_move_mapping(struct address_space *mapping, | |
405 | struct page *newpage, struct page *page) | |
406 | { | |
407 | int expected_count; | |
408 | void **pslot; | |
409 | ||
410 | if (!mapping) { | |
411 | if (page_count(page) != 1) | |
412 | return -EAGAIN; | |
78bd5209 | 413 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
414 | } |
415 | ||
416 | spin_lock_irq(&mapping->tree_lock); | |
417 | ||
418 | pslot = radix_tree_lookup_slot(&mapping->page_tree, | |
419 | page_index(page)); | |
420 | ||
421 | expected_count = 2 + page_has_private(page); | |
422 | if (page_count(page) != expected_count || | |
29c1f677 | 423 | radix_tree_deref_slot_protected(pslot, &mapping->tree_lock) != page) { |
290408d4 NH |
424 | spin_unlock_irq(&mapping->tree_lock); |
425 | return -EAGAIN; | |
426 | } | |
427 | ||
428 | if (!page_freeze_refs(page, expected_count)) { | |
429 | spin_unlock_irq(&mapping->tree_lock); | |
430 | return -EAGAIN; | |
431 | } | |
432 | ||
433 | get_page(newpage); | |
434 | ||
435 | radix_tree_replace_slot(pslot, newpage); | |
436 | ||
937a94c9 | 437 | page_unfreeze_refs(page, expected_count - 1); |
290408d4 NH |
438 | |
439 | spin_unlock_irq(&mapping->tree_lock); | |
78bd5209 | 440 | return MIGRATEPAGE_SUCCESS; |
290408d4 NH |
441 | } |
442 | ||
b20a3503 CL |
443 | /* |
444 | * Copy the page to its new location | |
445 | */ | |
290408d4 | 446 | void migrate_page_copy(struct page *newpage, struct page *page) |
b20a3503 | 447 | { |
7851a45c RR |
448 | int cpupid; |
449 | ||
b32967ff | 450 | if (PageHuge(page) || PageTransHuge(page)) |
290408d4 NH |
451 | copy_huge_page(newpage, page); |
452 | else | |
453 | copy_highpage(newpage, page); | |
b20a3503 CL |
454 | |
455 | if (PageError(page)) | |
456 | SetPageError(newpage); | |
457 | if (PageReferenced(page)) | |
458 | SetPageReferenced(newpage); | |
459 | if (PageUptodate(page)) | |
460 | SetPageUptodate(newpage); | |
894bc310 LS |
461 | if (TestClearPageActive(page)) { |
462 | VM_BUG_ON(PageUnevictable(page)); | |
b20a3503 | 463 | SetPageActive(newpage); |
418b27ef LS |
464 | } else if (TestClearPageUnevictable(page)) |
465 | SetPageUnevictable(newpage); | |
b20a3503 CL |
466 | if (PageChecked(page)) |
467 | SetPageChecked(newpage); | |
468 | if (PageMappedToDisk(page)) | |
469 | SetPageMappedToDisk(newpage); | |
470 | ||
471 | if (PageDirty(page)) { | |
472 | clear_page_dirty_for_io(page); | |
3a902c5f NP |
473 | /* |
474 | * Want to mark the page and the radix tree as dirty, and | |
475 | * redo the accounting that clear_page_dirty_for_io undid, | |
476 | * but we can't use set_page_dirty because that function | |
477 | * is actually a signal that all of the page has become dirty. | |
25985edc | 478 | * Whereas only part of our page may be dirty. |
3a902c5f | 479 | */ |
752dc185 HD |
480 | if (PageSwapBacked(page)) |
481 | SetPageDirty(newpage); | |
482 | else | |
483 | __set_page_dirty_nobuffers(newpage); | |
b20a3503 CL |
484 | } |
485 | ||
7851a45c RR |
486 | /* |
487 | * Copy NUMA information to the new page, to prevent over-eager | |
488 | * future migrations of this same page. | |
489 | */ | |
490 | cpupid = page_cpupid_xchg_last(page, -1); | |
491 | page_cpupid_xchg_last(newpage, cpupid); | |
492 | ||
b291f000 | 493 | mlock_migrate_page(newpage, page); |
e9995ef9 | 494 | ksm_migrate_page(newpage, page); |
c8d6553b HD |
495 | /* |
496 | * Please do not reorder this without considering how mm/ksm.c's | |
497 | * get_ksm_page() depends upon ksm_migrate_page() and PageSwapCache(). | |
498 | */ | |
b20a3503 | 499 | ClearPageSwapCache(page); |
b20a3503 CL |
500 | ClearPagePrivate(page); |
501 | set_page_private(page, 0); | |
b20a3503 CL |
502 | |
503 | /* | |
504 | * If any waiters have accumulated on the new page then | |
505 | * wake them up. | |
506 | */ | |
507 | if (PageWriteback(newpage)) | |
508 | end_page_writeback(newpage); | |
509 | } | |
b20a3503 | 510 | |
1d8b85cc CL |
511 | /************************************************************ |
512 | * Migration functions | |
513 | ***********************************************************/ | |
514 | ||
515 | /* Always fail migration. Used for mappings that are not movable */ | |
2d1db3b1 CL |
516 | int fail_migrate_page(struct address_space *mapping, |
517 | struct page *newpage, struct page *page) | |
1d8b85cc CL |
518 | { |
519 | return -EIO; | |
520 | } | |
521 | EXPORT_SYMBOL(fail_migrate_page); | |
522 | ||
b20a3503 CL |
523 | /* |
524 | * Common logic to directly migrate a single page suitable for | |
266cf658 | 525 | * pages that do not use PagePrivate/PagePrivate2. |
b20a3503 CL |
526 | * |
527 | * Pages are locked upon entry and exit. | |
528 | */ | |
2d1db3b1 | 529 | int migrate_page(struct address_space *mapping, |
a6bc32b8 MG |
530 | struct page *newpage, struct page *page, |
531 | enum migrate_mode mode) | |
b20a3503 CL |
532 | { |
533 | int rc; | |
534 | ||
535 | BUG_ON(PageWriteback(page)); /* Writeback must be complete */ | |
536 | ||
a6bc32b8 | 537 | rc = migrate_page_move_mapping(mapping, newpage, page, NULL, mode); |
b20a3503 | 538 | |
78bd5209 | 539 | if (rc != MIGRATEPAGE_SUCCESS) |
b20a3503 CL |
540 | return rc; |
541 | ||
542 | migrate_page_copy(newpage, page); | |
78bd5209 | 543 | return MIGRATEPAGE_SUCCESS; |
b20a3503 CL |
544 | } |
545 | EXPORT_SYMBOL(migrate_page); | |
546 | ||
9361401e | 547 | #ifdef CONFIG_BLOCK |
1d8b85cc CL |
548 | /* |
549 | * Migration function for pages with buffers. This function can only be used | |
550 | * if the underlying filesystem guarantees that no other references to "page" | |
551 | * exist. | |
552 | */ | |
2d1db3b1 | 553 | int buffer_migrate_page(struct address_space *mapping, |
a6bc32b8 | 554 | struct page *newpage, struct page *page, enum migrate_mode mode) |
1d8b85cc | 555 | { |
1d8b85cc CL |
556 | struct buffer_head *bh, *head; |
557 | int rc; | |
558 | ||
1d8b85cc | 559 | if (!page_has_buffers(page)) |
a6bc32b8 | 560 | return migrate_page(mapping, newpage, page, mode); |
1d8b85cc CL |
561 | |
562 | head = page_buffers(page); | |
563 | ||
a6bc32b8 | 564 | rc = migrate_page_move_mapping(mapping, newpage, page, head, mode); |
1d8b85cc | 565 | |
78bd5209 | 566 | if (rc != MIGRATEPAGE_SUCCESS) |
1d8b85cc CL |
567 | return rc; |
568 | ||
b969c4ab MG |
569 | /* |
570 | * In the async case, migrate_page_move_mapping locked the buffers | |
571 | * with an IRQ-safe spinlock held. In the sync case, the buffers | |
572 | * need to be locked now | |
573 | */ | |
a6bc32b8 MG |
574 | if (mode != MIGRATE_ASYNC) |
575 | BUG_ON(!buffer_migrate_lock_buffers(head, mode)); | |
1d8b85cc CL |
576 | |
577 | ClearPagePrivate(page); | |
578 | set_page_private(newpage, page_private(page)); | |
579 | set_page_private(page, 0); | |
580 | put_page(page); | |
581 | get_page(newpage); | |
582 | ||
583 | bh = head; | |
584 | do { | |
585 | set_bh_page(bh, newpage, bh_offset(bh)); | |
586 | bh = bh->b_this_page; | |
587 | ||
588 | } while (bh != head); | |
589 | ||
590 | SetPagePrivate(newpage); | |
591 | ||
592 | migrate_page_copy(newpage, page); | |
593 | ||
594 | bh = head; | |
595 | do { | |
596 | unlock_buffer(bh); | |
597 | put_bh(bh); | |
598 | bh = bh->b_this_page; | |
599 | ||
600 | } while (bh != head); | |
601 | ||
78bd5209 | 602 | return MIGRATEPAGE_SUCCESS; |
1d8b85cc CL |
603 | } |
604 | EXPORT_SYMBOL(buffer_migrate_page); | |
9361401e | 605 | #endif |
1d8b85cc | 606 | |
04e62a29 CL |
607 | /* |
608 | * Writeback a page to clean the dirty state | |
609 | */ | |
610 | static int writeout(struct address_space *mapping, struct page *page) | |
8351a6e4 | 611 | { |
04e62a29 CL |
612 | struct writeback_control wbc = { |
613 | .sync_mode = WB_SYNC_NONE, | |
614 | .nr_to_write = 1, | |
615 | .range_start = 0, | |
616 | .range_end = LLONG_MAX, | |
04e62a29 CL |
617 | .for_reclaim = 1 |
618 | }; | |
619 | int rc; | |
620 | ||
621 | if (!mapping->a_ops->writepage) | |
622 | /* No write method for the address space */ | |
623 | return -EINVAL; | |
624 | ||
625 | if (!clear_page_dirty_for_io(page)) | |
626 | /* Someone else already triggered a write */ | |
627 | return -EAGAIN; | |
628 | ||
8351a6e4 | 629 | /* |
04e62a29 CL |
630 | * A dirty page may imply that the underlying filesystem has |
631 | * the page on some queue. So the page must be clean for | |
632 | * migration. Writeout may mean we loose the lock and the | |
633 | * page state is no longer what we checked for earlier. | |
634 | * At this point we know that the migration attempt cannot | |
635 | * be successful. | |
8351a6e4 | 636 | */ |
04e62a29 | 637 | remove_migration_ptes(page, page); |
8351a6e4 | 638 | |
04e62a29 | 639 | rc = mapping->a_ops->writepage(page, &wbc); |
8351a6e4 | 640 | |
04e62a29 CL |
641 | if (rc != AOP_WRITEPAGE_ACTIVATE) |
642 | /* unlocked. Relock */ | |
643 | lock_page(page); | |
644 | ||
bda8550d | 645 | return (rc < 0) ? -EIO : -EAGAIN; |
04e62a29 CL |
646 | } |
647 | ||
648 | /* | |
649 | * Default handling if a filesystem does not provide a migration function. | |
650 | */ | |
651 | static int fallback_migrate_page(struct address_space *mapping, | |
a6bc32b8 | 652 | struct page *newpage, struct page *page, enum migrate_mode mode) |
04e62a29 | 653 | { |
b969c4ab | 654 | if (PageDirty(page)) { |
a6bc32b8 MG |
655 | /* Only writeback pages in full synchronous migration */ |
656 | if (mode != MIGRATE_SYNC) | |
b969c4ab | 657 | return -EBUSY; |
04e62a29 | 658 | return writeout(mapping, page); |
b969c4ab | 659 | } |
8351a6e4 CL |
660 | |
661 | /* | |
662 | * Buffers may be managed in a filesystem specific way. | |
663 | * We must have no buffers or drop them. | |
664 | */ | |
266cf658 | 665 | if (page_has_private(page) && |
8351a6e4 CL |
666 | !try_to_release_page(page, GFP_KERNEL)) |
667 | return -EAGAIN; | |
668 | ||
a6bc32b8 | 669 | return migrate_page(mapping, newpage, page, mode); |
8351a6e4 CL |
670 | } |
671 | ||
e24f0b8f CL |
672 | /* |
673 | * Move a page to a newly allocated page | |
674 | * The page is locked and all ptes have been successfully removed. | |
675 | * | |
676 | * The new page will have replaced the old page if this function | |
677 | * is successful. | |
894bc310 LS |
678 | * |
679 | * Return value: | |
680 | * < 0 - error code | |
78bd5209 | 681 | * MIGRATEPAGE_SUCCESS - success |
e24f0b8f | 682 | */ |
3fe2011f | 683 | static int move_to_new_page(struct page *newpage, struct page *page, |
a6bc32b8 | 684 | int remap_swapcache, enum migrate_mode mode) |
e24f0b8f CL |
685 | { |
686 | struct address_space *mapping; | |
687 | int rc; | |
688 | ||
689 | /* | |
690 | * Block others from accessing the page when we get around to | |
691 | * establishing additional references. We are the only one | |
692 | * holding a reference to the new page at this point. | |
693 | */ | |
529ae9aa | 694 | if (!trylock_page(newpage)) |
e24f0b8f CL |
695 | BUG(); |
696 | ||
697 | /* Prepare mapping for the new page.*/ | |
698 | newpage->index = page->index; | |
699 | newpage->mapping = page->mapping; | |
b2e18538 RR |
700 | if (PageSwapBacked(page)) |
701 | SetPageSwapBacked(newpage); | |
e24f0b8f CL |
702 | |
703 | mapping = page_mapping(page); | |
704 | if (!mapping) | |
a6bc32b8 | 705 | rc = migrate_page(mapping, newpage, page, mode); |
b969c4ab | 706 | else if (mapping->a_ops->migratepage) |
e24f0b8f | 707 | /* |
b969c4ab MG |
708 | * Most pages have a mapping and most filesystems provide a |
709 | * migratepage callback. Anonymous pages are part of swap | |
710 | * space which also has its own migratepage callback. This | |
711 | * is the most common path for page migration. | |
e24f0b8f | 712 | */ |
b969c4ab | 713 | rc = mapping->a_ops->migratepage(mapping, |
a6bc32b8 | 714 | newpage, page, mode); |
b969c4ab | 715 | else |
a6bc32b8 | 716 | rc = fallback_migrate_page(mapping, newpage, page, mode); |
e24f0b8f | 717 | |
78bd5209 | 718 | if (rc != MIGRATEPAGE_SUCCESS) { |
e24f0b8f | 719 | newpage->mapping = NULL; |
3fe2011f MG |
720 | } else { |
721 | if (remap_swapcache) | |
722 | remove_migration_ptes(page, newpage); | |
35512eca | 723 | page->mapping = NULL; |
3fe2011f | 724 | } |
e24f0b8f CL |
725 | |
726 | unlock_page(newpage); | |
727 | ||
728 | return rc; | |
729 | } | |
730 | ||
0dabec93 | 731 | static int __unmap_and_move(struct page *page, struct page *newpage, |
9c620e2b | 732 | int force, enum migrate_mode mode) |
e24f0b8f | 733 | { |
0dabec93 | 734 | int rc = -EAGAIN; |
3fe2011f | 735 | int remap_swapcache = 1; |
56039efa | 736 | struct mem_cgroup *mem; |
3f6c8272 | 737 | struct anon_vma *anon_vma = NULL; |
95a402c3 | 738 | |
529ae9aa | 739 | if (!trylock_page(page)) { |
a6bc32b8 | 740 | if (!force || mode == MIGRATE_ASYNC) |
0dabec93 | 741 | goto out; |
3e7d3449 MG |
742 | |
743 | /* | |
744 | * It's not safe for direct compaction to call lock_page. | |
745 | * For example, during page readahead pages are added locked | |
746 | * to the LRU. Later, when the IO completes the pages are | |
747 | * marked uptodate and unlocked. However, the queueing | |
748 | * could be merging multiple pages for one bio (e.g. | |
749 | * mpage_readpages). If an allocation happens for the | |
750 | * second or third page, the process can end up locking | |
751 | * the same page twice and deadlocking. Rather than | |
752 | * trying to be clever about what pages can be locked, | |
753 | * avoid the use of lock_page for direct compaction | |
754 | * altogether. | |
755 | */ | |
756 | if (current->flags & PF_MEMALLOC) | |
0dabec93 | 757 | goto out; |
3e7d3449 | 758 | |
e24f0b8f CL |
759 | lock_page(page); |
760 | } | |
761 | ||
01b1ae63 | 762 | /* charge against new page */ |
0030f535 | 763 | mem_cgroup_prepare_migration(page, newpage, &mem); |
01b1ae63 | 764 | |
e24f0b8f | 765 | if (PageWriteback(page)) { |
11bc82d6 | 766 | /* |
fed5b64a | 767 | * Only in the case of a full synchronous migration is it |
a6bc32b8 MG |
768 | * necessary to wait for PageWriteback. In the async case, |
769 | * the retry loop is too short and in the sync-light case, | |
770 | * the overhead of stalling is too much | |
11bc82d6 | 771 | */ |
a6bc32b8 | 772 | if (mode != MIGRATE_SYNC) { |
11bc82d6 AA |
773 | rc = -EBUSY; |
774 | goto uncharge; | |
775 | } | |
776 | if (!force) | |
01b1ae63 | 777 | goto uncharge; |
e24f0b8f CL |
778 | wait_on_page_writeback(page); |
779 | } | |
e24f0b8f | 780 | /* |
dc386d4d KH |
781 | * By try_to_unmap(), page->mapcount goes down to 0 here. In this case, |
782 | * we cannot notice that anon_vma is freed while we migrates a page. | |
1ce82b69 | 783 | * This get_anon_vma() delays freeing anon_vma pointer until the end |
dc386d4d | 784 | * of migration. File cache pages are no problem because of page_lock() |
989f89c5 KH |
785 | * File Caches may use write_page() or lock_page() in migration, then, |
786 | * just care Anon page here. | |
dc386d4d | 787 | */ |
b79bc0a0 | 788 | if (PageAnon(page) && !PageKsm(page)) { |
1ce82b69 | 789 | /* |
4fc3f1d6 | 790 | * Only page_lock_anon_vma_read() understands the subtleties of |
1ce82b69 HD |
791 | * getting a hold on an anon_vma from outside one of its mms. |
792 | */ | |
746b18d4 | 793 | anon_vma = page_get_anon_vma(page); |
1ce82b69 HD |
794 | if (anon_vma) { |
795 | /* | |
746b18d4 | 796 | * Anon page |
1ce82b69 | 797 | */ |
1ce82b69 | 798 | } else if (PageSwapCache(page)) { |
3fe2011f MG |
799 | /* |
800 | * We cannot be sure that the anon_vma of an unmapped | |
801 | * swapcache page is safe to use because we don't | |
802 | * know in advance if the VMA that this page belonged | |
803 | * to still exists. If the VMA and others sharing the | |
804 | * data have been freed, then the anon_vma could | |
805 | * already be invalid. | |
806 | * | |
807 | * To avoid this possibility, swapcache pages get | |
808 | * migrated but are not remapped when migration | |
809 | * completes | |
810 | */ | |
811 | remap_swapcache = 0; | |
812 | } else { | |
1ce82b69 | 813 | goto uncharge; |
3fe2011f | 814 | } |
989f89c5 | 815 | } |
62e1c553 | 816 | |
bf6bddf1 RA |
817 | if (unlikely(balloon_page_movable(page))) { |
818 | /* | |
819 | * A ballooned page does not need any special attention from | |
820 | * physical to virtual reverse mapping procedures. | |
821 | * Skip any attempt to unmap PTEs or to remap swap cache, | |
822 | * in order to avoid burning cycles at rmap level, and perform | |
823 | * the page migration right away (proteced by page lock). | |
824 | */ | |
825 | rc = balloon_page_migrate(newpage, page, mode); | |
826 | goto uncharge; | |
827 | } | |
828 | ||
dc386d4d | 829 | /* |
62e1c553 SL |
830 | * Corner case handling: |
831 | * 1. When a new swap-cache page is read into, it is added to the LRU | |
832 | * and treated as swapcache but it has no rmap yet. | |
833 | * Calling try_to_unmap() against a page->mapping==NULL page will | |
834 | * trigger a BUG. So handle it here. | |
835 | * 2. An orphaned page (see truncate_complete_page) might have | |
836 | * fs-private metadata. The page can be picked up due to memory | |
837 | * offlining. Everywhere else except page reclaim, the page is | |
838 | * invisible to the vm, so the page can not be migrated. So try to | |
839 | * free the metadata, so the page can be freed. | |
e24f0b8f | 840 | */ |
62e1c553 | 841 | if (!page->mapping) { |
1ce82b69 HD |
842 | VM_BUG_ON(PageAnon(page)); |
843 | if (page_has_private(page)) { | |
62e1c553 | 844 | try_to_free_buffers(page); |
1ce82b69 | 845 | goto uncharge; |
62e1c553 | 846 | } |
abfc3488 | 847 | goto skip_unmap; |
62e1c553 SL |
848 | } |
849 | ||
dc386d4d | 850 | /* Establish migration ptes or remove ptes */ |
14fa31b8 | 851 | try_to_unmap(page, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); |
dc386d4d | 852 | |
abfc3488 | 853 | skip_unmap: |
e6a1530d | 854 | if (!page_mapped(page)) |
a6bc32b8 | 855 | rc = move_to_new_page(newpage, page, remap_swapcache, mode); |
e24f0b8f | 856 | |
3fe2011f | 857 | if (rc && remap_swapcache) |
e24f0b8f | 858 | remove_migration_ptes(page, page); |
3f6c8272 MG |
859 | |
860 | /* Drop an anon_vma reference if we took one */ | |
76545066 | 861 | if (anon_vma) |
9e60109f | 862 | put_anon_vma(anon_vma); |
3f6c8272 | 863 | |
01b1ae63 | 864 | uncharge: |
bf6bddf1 RA |
865 | mem_cgroup_end_migration(mem, page, newpage, |
866 | (rc == MIGRATEPAGE_SUCCESS || | |
867 | rc == MIGRATEPAGE_BALLOON_SUCCESS)); | |
e24f0b8f | 868 | unlock_page(page); |
0dabec93 MK |
869 | out: |
870 | return rc; | |
871 | } | |
95a402c3 | 872 | |
0dabec93 MK |
873 | /* |
874 | * Obtain the lock on page, remove all ptes and migrate the page | |
875 | * to the newly allocated page in newpage. | |
876 | */ | |
877 | static int unmap_and_move(new_page_t get_new_page, unsigned long private, | |
9c620e2b | 878 | struct page *page, int force, enum migrate_mode mode) |
0dabec93 MK |
879 | { |
880 | int rc = 0; | |
881 | int *result = NULL; | |
882 | struct page *newpage = get_new_page(page, private, &result); | |
883 | ||
884 | if (!newpage) | |
885 | return -ENOMEM; | |
886 | ||
887 | if (page_count(page) == 1) { | |
888 | /* page was freed from under us. So we are done. */ | |
889 | goto out; | |
890 | } | |
891 | ||
892 | if (unlikely(PageTransHuge(page))) | |
893 | if (unlikely(split_huge_page(page))) | |
894 | goto out; | |
895 | ||
9c620e2b | 896 | rc = __unmap_and_move(page, newpage, force, mode); |
bf6bddf1 RA |
897 | |
898 | if (unlikely(rc == MIGRATEPAGE_BALLOON_SUCCESS)) { | |
899 | /* | |
900 | * A ballooned page has been migrated already. | |
901 | * Now, it's the time to wrap-up counters, | |
902 | * handle the page back to Buddy and return. | |
903 | */ | |
904 | dec_zone_page_state(page, NR_ISOLATED_ANON + | |
905 | page_is_file_cache(page)); | |
906 | balloon_page_free(page); | |
907 | return MIGRATEPAGE_SUCCESS; | |
908 | } | |
0dabec93 | 909 | out: |
e24f0b8f | 910 | if (rc != -EAGAIN) { |
0dabec93 MK |
911 | /* |
912 | * A page that has been migrated has all references | |
913 | * removed and will be freed. A page that has not been | |
914 | * migrated will have kepts its references and be | |
915 | * restored. | |
916 | */ | |
917 | list_del(&page->lru); | |
a731286d | 918 | dec_zone_page_state(page, NR_ISOLATED_ANON + |
6c0b1351 | 919 | page_is_file_cache(page)); |
894bc310 | 920 | putback_lru_page(page); |
e24f0b8f | 921 | } |
95a402c3 CL |
922 | /* |
923 | * Move the new page to the LRU. If migration was not successful | |
924 | * then this will free the page. | |
925 | */ | |
894bc310 | 926 | putback_lru_page(newpage); |
742755a1 CL |
927 | if (result) { |
928 | if (rc) | |
929 | *result = rc; | |
930 | else | |
931 | *result = page_to_nid(newpage); | |
932 | } | |
e24f0b8f CL |
933 | return rc; |
934 | } | |
935 | ||
290408d4 NH |
936 | /* |
937 | * Counterpart of unmap_and_move_page() for hugepage migration. | |
938 | * | |
939 | * This function doesn't wait the completion of hugepage I/O | |
940 | * because there is no race between I/O and migration for hugepage. | |
941 | * Note that currently hugepage I/O occurs only in direct I/O | |
942 | * where no lock is held and PG_writeback is irrelevant, | |
943 | * and writeback status of all subpages are counted in the reference | |
944 | * count of the head page (i.e. if all subpages of a 2MB hugepage are | |
945 | * under direct I/O, the reference of the head page is 512 and a bit more.) | |
946 | * This means that when we try to migrate hugepage whose subpages are | |
947 | * doing direct I/O, some references remain after try_to_unmap() and | |
948 | * hugepage migration fails without data corruption. | |
949 | * | |
950 | * There is also no race when direct I/O is issued on the page under migration, | |
951 | * because then pte is replaced with migration swap entry and direct I/O code | |
952 | * will wait in the page fault for migration to complete. | |
953 | */ | |
954 | static int unmap_and_move_huge_page(new_page_t get_new_page, | |
955 | unsigned long private, struct page *hpage, | |
9c620e2b | 956 | int force, enum migrate_mode mode) |
290408d4 NH |
957 | { |
958 | int rc = 0; | |
959 | int *result = NULL; | |
960 | struct page *new_hpage = get_new_page(hpage, private, &result); | |
290408d4 NH |
961 | struct anon_vma *anon_vma = NULL; |
962 | ||
83467efb NH |
963 | /* |
964 | * Movability of hugepages depends on architectures and hugepage size. | |
965 | * This check is necessary because some callers of hugepage migration | |
966 | * like soft offline and memory hotremove don't walk through page | |
967 | * tables or check whether the hugepage is pmd-based or not before | |
968 | * kicking migration. | |
969 | */ | |
970 | if (!hugepage_migration_support(page_hstate(hpage))) | |
971 | return -ENOSYS; | |
972 | ||
290408d4 NH |
973 | if (!new_hpage) |
974 | return -ENOMEM; | |
975 | ||
976 | rc = -EAGAIN; | |
977 | ||
978 | if (!trylock_page(hpage)) { | |
a6bc32b8 | 979 | if (!force || mode != MIGRATE_SYNC) |
290408d4 NH |
980 | goto out; |
981 | lock_page(hpage); | |
982 | } | |
983 | ||
746b18d4 PZ |
984 | if (PageAnon(hpage)) |
985 | anon_vma = page_get_anon_vma(hpage); | |
290408d4 NH |
986 | |
987 | try_to_unmap(hpage, TTU_MIGRATION|TTU_IGNORE_MLOCK|TTU_IGNORE_ACCESS); | |
988 | ||
989 | if (!page_mapped(hpage)) | |
a6bc32b8 | 990 | rc = move_to_new_page(new_hpage, hpage, 1, mode); |
290408d4 NH |
991 | |
992 | if (rc) | |
993 | remove_migration_ptes(hpage, hpage); | |
994 | ||
fd4a4663 | 995 | if (anon_vma) |
9e60109f | 996 | put_anon_vma(anon_vma); |
8e6ac7fa AK |
997 | |
998 | if (!rc) | |
999 | hugetlb_cgroup_migrate(hpage, new_hpage); | |
1000 | ||
290408d4 | 1001 | unlock_page(hpage); |
09761333 | 1002 | out: |
b8ec1cee NH |
1003 | if (rc != -EAGAIN) |
1004 | putback_active_hugepage(hpage); | |
290408d4 | 1005 | put_page(new_hpage); |
290408d4 NH |
1006 | if (result) { |
1007 | if (rc) | |
1008 | *result = rc; | |
1009 | else | |
1010 | *result = page_to_nid(new_hpage); | |
1011 | } | |
1012 | return rc; | |
1013 | } | |
1014 | ||
b20a3503 | 1015 | /* |
c73e5c9c SB |
1016 | * migrate_pages - migrate the pages specified in a list, to the free pages |
1017 | * supplied as the target for the page migration | |
b20a3503 | 1018 | * |
c73e5c9c SB |
1019 | * @from: The list of pages to be migrated. |
1020 | * @get_new_page: The function used to allocate free pages to be used | |
1021 | * as the target of the page migration. | |
1022 | * @private: Private data to be passed on to get_new_page() | |
1023 | * @mode: The migration mode that specifies the constraints for | |
1024 | * page migration, if any. | |
1025 | * @reason: The reason for page migration. | |
b20a3503 | 1026 | * |
c73e5c9c SB |
1027 | * The function returns after 10 attempts or if no pages are movable any more |
1028 | * because the list has become empty or no retryable pages exist any more. | |
1029 | * The caller should call putback_lru_pages() to return pages to the LRU | |
28bd6578 | 1030 | * or free list only if ret != 0. |
b20a3503 | 1031 | * |
c73e5c9c | 1032 | * Returns the number of pages that were not migrated, or an error code. |
b20a3503 | 1033 | */ |
9c620e2b HD |
1034 | int migrate_pages(struct list_head *from, new_page_t get_new_page, |
1035 | unsigned long private, enum migrate_mode mode, int reason) | |
b20a3503 | 1036 | { |
e24f0b8f | 1037 | int retry = 1; |
b20a3503 | 1038 | int nr_failed = 0; |
5647bc29 | 1039 | int nr_succeeded = 0; |
b20a3503 CL |
1040 | int pass = 0; |
1041 | struct page *page; | |
1042 | struct page *page2; | |
1043 | int swapwrite = current->flags & PF_SWAPWRITE; | |
1044 | int rc; | |
1045 | ||
1046 | if (!swapwrite) | |
1047 | current->flags |= PF_SWAPWRITE; | |
1048 | ||
e24f0b8f CL |
1049 | for(pass = 0; pass < 10 && retry; pass++) { |
1050 | retry = 0; | |
b20a3503 | 1051 | |
e24f0b8f | 1052 | list_for_each_entry_safe(page, page2, from, lru) { |
e24f0b8f | 1053 | cond_resched(); |
2d1db3b1 | 1054 | |
31caf665 NH |
1055 | if (PageHuge(page)) |
1056 | rc = unmap_and_move_huge_page(get_new_page, | |
1057 | private, page, pass > 2, mode); | |
1058 | else | |
1059 | rc = unmap_and_move(get_new_page, private, | |
9c620e2b | 1060 | page, pass > 2, mode); |
2d1db3b1 | 1061 | |
e24f0b8f | 1062 | switch(rc) { |
95a402c3 CL |
1063 | case -ENOMEM: |
1064 | goto out; | |
e24f0b8f | 1065 | case -EAGAIN: |
2d1db3b1 | 1066 | retry++; |
e24f0b8f | 1067 | break; |
78bd5209 | 1068 | case MIGRATEPAGE_SUCCESS: |
5647bc29 | 1069 | nr_succeeded++; |
e24f0b8f CL |
1070 | break; |
1071 | default: | |
2d1db3b1 | 1072 | /* Permanent failure */ |
2d1db3b1 | 1073 | nr_failed++; |
e24f0b8f | 1074 | break; |
2d1db3b1 | 1075 | } |
b20a3503 CL |
1076 | } |
1077 | } | |
78bd5209 | 1078 | rc = nr_failed + retry; |
95a402c3 | 1079 | out: |
5647bc29 MG |
1080 | if (nr_succeeded) |
1081 | count_vm_events(PGMIGRATE_SUCCESS, nr_succeeded); | |
1082 | if (nr_failed) | |
1083 | count_vm_events(PGMIGRATE_FAIL, nr_failed); | |
7b2a2d4a MG |
1084 | trace_mm_migrate_pages(nr_succeeded, nr_failed, mode, reason); |
1085 | ||
b20a3503 CL |
1086 | if (!swapwrite) |
1087 | current->flags &= ~PF_SWAPWRITE; | |
1088 | ||
78bd5209 | 1089 | return rc; |
b20a3503 | 1090 | } |
95a402c3 | 1091 | |
742755a1 CL |
1092 | #ifdef CONFIG_NUMA |
1093 | /* | |
1094 | * Move a list of individual pages | |
1095 | */ | |
1096 | struct page_to_node { | |
1097 | unsigned long addr; | |
1098 | struct page *page; | |
1099 | int node; | |
1100 | int status; | |
1101 | }; | |
1102 | ||
1103 | static struct page *new_page_node(struct page *p, unsigned long private, | |
1104 | int **result) | |
1105 | { | |
1106 | struct page_to_node *pm = (struct page_to_node *)private; | |
1107 | ||
1108 | while (pm->node != MAX_NUMNODES && pm->page != p) | |
1109 | pm++; | |
1110 | ||
1111 | if (pm->node == MAX_NUMNODES) | |
1112 | return NULL; | |
1113 | ||
1114 | *result = &pm->status; | |
1115 | ||
e632a938 NH |
1116 | if (PageHuge(p)) |
1117 | return alloc_huge_page_node(page_hstate(compound_head(p)), | |
1118 | pm->node); | |
1119 | else | |
1120 | return alloc_pages_exact_node(pm->node, | |
769848c0 | 1121 | GFP_HIGHUSER_MOVABLE | GFP_THISNODE, 0); |
742755a1 CL |
1122 | } |
1123 | ||
1124 | /* | |
1125 | * Move a set of pages as indicated in the pm array. The addr | |
1126 | * field must be set to the virtual address of the page to be moved | |
1127 | * and the node number must contain a valid target node. | |
5e9a0f02 | 1128 | * The pm array ends with node = MAX_NUMNODES. |
742755a1 | 1129 | */ |
5e9a0f02 BG |
1130 | static int do_move_page_to_node_array(struct mm_struct *mm, |
1131 | struct page_to_node *pm, | |
1132 | int migrate_all) | |
742755a1 CL |
1133 | { |
1134 | int err; | |
1135 | struct page_to_node *pp; | |
1136 | LIST_HEAD(pagelist); | |
1137 | ||
1138 | down_read(&mm->mmap_sem); | |
1139 | ||
1140 | /* | |
1141 | * Build a list of pages to migrate | |
1142 | */ | |
742755a1 CL |
1143 | for (pp = pm; pp->node != MAX_NUMNODES; pp++) { |
1144 | struct vm_area_struct *vma; | |
1145 | struct page *page; | |
1146 | ||
742755a1 CL |
1147 | err = -EFAULT; |
1148 | vma = find_vma(mm, pp->addr); | |
70384dc6 | 1149 | if (!vma || pp->addr < vma->vm_start || !vma_migratable(vma)) |
742755a1 CL |
1150 | goto set_status; |
1151 | ||
500d65d4 | 1152 | page = follow_page(vma, pp->addr, FOLL_GET|FOLL_SPLIT); |
89f5b7da LT |
1153 | |
1154 | err = PTR_ERR(page); | |
1155 | if (IS_ERR(page)) | |
1156 | goto set_status; | |
1157 | ||
742755a1 CL |
1158 | err = -ENOENT; |
1159 | if (!page) | |
1160 | goto set_status; | |
1161 | ||
62b61f61 | 1162 | /* Use PageReserved to check for zero page */ |
b79bc0a0 | 1163 | if (PageReserved(page)) |
742755a1 CL |
1164 | goto put_and_set; |
1165 | ||
1166 | pp->page = page; | |
1167 | err = page_to_nid(page); | |
1168 | ||
1169 | if (err == pp->node) | |
1170 | /* | |
1171 | * Node already in the right place | |
1172 | */ | |
1173 | goto put_and_set; | |
1174 | ||
1175 | err = -EACCES; | |
1176 | if (page_mapcount(page) > 1 && | |
1177 | !migrate_all) | |
1178 | goto put_and_set; | |
1179 | ||
e632a938 NH |
1180 | if (PageHuge(page)) { |
1181 | isolate_huge_page(page, &pagelist); | |
1182 | goto put_and_set; | |
1183 | } | |
1184 | ||
62695a84 | 1185 | err = isolate_lru_page(page); |
6d9c285a | 1186 | if (!err) { |
62695a84 | 1187 | list_add_tail(&page->lru, &pagelist); |
6d9c285a KM |
1188 | inc_zone_page_state(page, NR_ISOLATED_ANON + |
1189 | page_is_file_cache(page)); | |
1190 | } | |
742755a1 CL |
1191 | put_and_set: |
1192 | /* | |
1193 | * Either remove the duplicate refcount from | |
1194 | * isolate_lru_page() or drop the page ref if it was | |
1195 | * not isolated. | |
1196 | */ | |
1197 | put_page(page); | |
1198 | set_status: | |
1199 | pp->status = err; | |
1200 | } | |
1201 | ||
e78bbfa8 | 1202 | err = 0; |
cf608ac1 | 1203 | if (!list_empty(&pagelist)) { |
742755a1 | 1204 | err = migrate_pages(&pagelist, new_page_node, |
9c620e2b | 1205 | (unsigned long)pm, MIGRATE_SYNC, MR_SYSCALL); |
cf608ac1 | 1206 | if (err) |
e632a938 | 1207 | putback_movable_pages(&pagelist); |
cf608ac1 | 1208 | } |
742755a1 CL |
1209 | |
1210 | up_read(&mm->mmap_sem); | |
1211 | return err; | |
1212 | } | |
1213 | ||
5e9a0f02 BG |
1214 | /* |
1215 | * Migrate an array of page address onto an array of nodes and fill | |
1216 | * the corresponding array of status. | |
1217 | */ | |
3268c63e | 1218 | static int do_pages_move(struct mm_struct *mm, nodemask_t task_nodes, |
5e9a0f02 BG |
1219 | unsigned long nr_pages, |
1220 | const void __user * __user *pages, | |
1221 | const int __user *nodes, | |
1222 | int __user *status, int flags) | |
1223 | { | |
3140a227 | 1224 | struct page_to_node *pm; |
3140a227 BG |
1225 | unsigned long chunk_nr_pages; |
1226 | unsigned long chunk_start; | |
1227 | int err; | |
5e9a0f02 | 1228 | |
3140a227 BG |
1229 | err = -ENOMEM; |
1230 | pm = (struct page_to_node *)__get_free_page(GFP_KERNEL); | |
1231 | if (!pm) | |
5e9a0f02 | 1232 | goto out; |
35282a2d BG |
1233 | |
1234 | migrate_prep(); | |
1235 | ||
5e9a0f02 | 1236 | /* |
3140a227 BG |
1237 | * Store a chunk of page_to_node array in a page, |
1238 | * but keep the last one as a marker | |
5e9a0f02 | 1239 | */ |
3140a227 | 1240 | chunk_nr_pages = (PAGE_SIZE / sizeof(struct page_to_node)) - 1; |
5e9a0f02 | 1241 | |
3140a227 BG |
1242 | for (chunk_start = 0; |
1243 | chunk_start < nr_pages; | |
1244 | chunk_start += chunk_nr_pages) { | |
1245 | int j; | |
5e9a0f02 | 1246 | |
3140a227 BG |
1247 | if (chunk_start + chunk_nr_pages > nr_pages) |
1248 | chunk_nr_pages = nr_pages - chunk_start; | |
1249 | ||
1250 | /* fill the chunk pm with addrs and nodes from user-space */ | |
1251 | for (j = 0; j < chunk_nr_pages; j++) { | |
1252 | const void __user *p; | |
5e9a0f02 BG |
1253 | int node; |
1254 | ||
3140a227 BG |
1255 | err = -EFAULT; |
1256 | if (get_user(p, pages + j + chunk_start)) | |
1257 | goto out_pm; | |
1258 | pm[j].addr = (unsigned long) p; | |
1259 | ||
1260 | if (get_user(node, nodes + j + chunk_start)) | |
5e9a0f02 BG |
1261 | goto out_pm; |
1262 | ||
1263 | err = -ENODEV; | |
6f5a55f1 LT |
1264 | if (node < 0 || node >= MAX_NUMNODES) |
1265 | goto out_pm; | |
1266 | ||
389162c2 | 1267 | if (!node_state(node, N_MEMORY)) |
5e9a0f02 BG |
1268 | goto out_pm; |
1269 | ||
1270 | err = -EACCES; | |
1271 | if (!node_isset(node, task_nodes)) | |
1272 | goto out_pm; | |
1273 | ||
3140a227 BG |
1274 | pm[j].node = node; |
1275 | } | |
1276 | ||
1277 | /* End marker for this chunk */ | |
1278 | pm[chunk_nr_pages].node = MAX_NUMNODES; | |
1279 | ||
1280 | /* Migrate this chunk */ | |
1281 | err = do_move_page_to_node_array(mm, pm, | |
1282 | flags & MPOL_MF_MOVE_ALL); | |
1283 | if (err < 0) | |
1284 | goto out_pm; | |
5e9a0f02 | 1285 | |
5e9a0f02 | 1286 | /* Return status information */ |
3140a227 BG |
1287 | for (j = 0; j < chunk_nr_pages; j++) |
1288 | if (put_user(pm[j].status, status + j + chunk_start)) { | |
5e9a0f02 | 1289 | err = -EFAULT; |
3140a227 BG |
1290 | goto out_pm; |
1291 | } | |
1292 | } | |
1293 | err = 0; | |
5e9a0f02 BG |
1294 | |
1295 | out_pm: | |
3140a227 | 1296 | free_page((unsigned long)pm); |
5e9a0f02 BG |
1297 | out: |
1298 | return err; | |
1299 | } | |
1300 | ||
742755a1 | 1301 | /* |
2f007e74 | 1302 | * Determine the nodes of an array of pages and store it in an array of status. |
742755a1 | 1303 | */ |
80bba129 BG |
1304 | static void do_pages_stat_array(struct mm_struct *mm, unsigned long nr_pages, |
1305 | const void __user **pages, int *status) | |
742755a1 | 1306 | { |
2f007e74 | 1307 | unsigned long i; |
2f007e74 | 1308 | |
742755a1 CL |
1309 | down_read(&mm->mmap_sem); |
1310 | ||
2f007e74 | 1311 | for (i = 0; i < nr_pages; i++) { |
80bba129 | 1312 | unsigned long addr = (unsigned long)(*pages); |
742755a1 CL |
1313 | struct vm_area_struct *vma; |
1314 | struct page *page; | |
c095adbc | 1315 | int err = -EFAULT; |
2f007e74 BG |
1316 | |
1317 | vma = find_vma(mm, addr); | |
70384dc6 | 1318 | if (!vma || addr < vma->vm_start) |
742755a1 CL |
1319 | goto set_status; |
1320 | ||
2f007e74 | 1321 | page = follow_page(vma, addr, 0); |
89f5b7da LT |
1322 | |
1323 | err = PTR_ERR(page); | |
1324 | if (IS_ERR(page)) | |
1325 | goto set_status; | |
1326 | ||
742755a1 CL |
1327 | err = -ENOENT; |
1328 | /* Use PageReserved to check for zero page */ | |
b79bc0a0 | 1329 | if (!page || PageReserved(page)) |
742755a1 CL |
1330 | goto set_status; |
1331 | ||
1332 | err = page_to_nid(page); | |
1333 | set_status: | |
80bba129 BG |
1334 | *status = err; |
1335 | ||
1336 | pages++; | |
1337 | status++; | |
1338 | } | |
1339 | ||
1340 | up_read(&mm->mmap_sem); | |
1341 | } | |
1342 | ||
1343 | /* | |
1344 | * Determine the nodes of a user array of pages and store it in | |
1345 | * a user array of status. | |
1346 | */ | |
1347 | static int do_pages_stat(struct mm_struct *mm, unsigned long nr_pages, | |
1348 | const void __user * __user *pages, | |
1349 | int __user *status) | |
1350 | { | |
1351 | #define DO_PAGES_STAT_CHUNK_NR 16 | |
1352 | const void __user *chunk_pages[DO_PAGES_STAT_CHUNK_NR]; | |
1353 | int chunk_status[DO_PAGES_STAT_CHUNK_NR]; | |
80bba129 | 1354 | |
87b8d1ad PA |
1355 | while (nr_pages) { |
1356 | unsigned long chunk_nr; | |
80bba129 | 1357 | |
87b8d1ad PA |
1358 | chunk_nr = nr_pages; |
1359 | if (chunk_nr > DO_PAGES_STAT_CHUNK_NR) | |
1360 | chunk_nr = DO_PAGES_STAT_CHUNK_NR; | |
1361 | ||
1362 | if (copy_from_user(chunk_pages, pages, chunk_nr * sizeof(*chunk_pages))) | |
1363 | break; | |
80bba129 BG |
1364 | |
1365 | do_pages_stat_array(mm, chunk_nr, chunk_pages, chunk_status); | |
1366 | ||
87b8d1ad PA |
1367 | if (copy_to_user(status, chunk_status, chunk_nr * sizeof(*status))) |
1368 | break; | |
742755a1 | 1369 | |
87b8d1ad PA |
1370 | pages += chunk_nr; |
1371 | status += chunk_nr; | |
1372 | nr_pages -= chunk_nr; | |
1373 | } | |
1374 | return nr_pages ? -EFAULT : 0; | |
742755a1 CL |
1375 | } |
1376 | ||
1377 | /* | |
1378 | * Move a list of pages in the address space of the currently executing | |
1379 | * process. | |
1380 | */ | |
938bb9f5 HC |
1381 | SYSCALL_DEFINE6(move_pages, pid_t, pid, unsigned long, nr_pages, |
1382 | const void __user * __user *, pages, | |
1383 | const int __user *, nodes, | |
1384 | int __user *, status, int, flags) | |
742755a1 | 1385 | { |
c69e8d9c | 1386 | const struct cred *cred = current_cred(), *tcred; |
742755a1 | 1387 | struct task_struct *task; |
742755a1 | 1388 | struct mm_struct *mm; |
5e9a0f02 | 1389 | int err; |
3268c63e | 1390 | nodemask_t task_nodes; |
742755a1 CL |
1391 | |
1392 | /* Check flags */ | |
1393 | if (flags & ~(MPOL_MF_MOVE|MPOL_MF_MOVE_ALL)) | |
1394 | return -EINVAL; | |
1395 | ||
1396 | if ((flags & MPOL_MF_MOVE_ALL) && !capable(CAP_SYS_NICE)) | |
1397 | return -EPERM; | |
1398 | ||
1399 | /* Find the mm_struct */ | |
a879bf58 | 1400 | rcu_read_lock(); |
228ebcbe | 1401 | task = pid ? find_task_by_vpid(pid) : current; |
742755a1 | 1402 | if (!task) { |
a879bf58 | 1403 | rcu_read_unlock(); |
742755a1 CL |
1404 | return -ESRCH; |
1405 | } | |
3268c63e | 1406 | get_task_struct(task); |
742755a1 CL |
1407 | |
1408 | /* | |
1409 | * Check if this process has the right to modify the specified | |
1410 | * process. The right exists if the process has administrative | |
1411 | * capabilities, superuser privileges or the same | |
1412 | * userid as the target process. | |
1413 | */ | |
c69e8d9c | 1414 | tcred = __task_cred(task); |
b38a86eb EB |
1415 | if (!uid_eq(cred->euid, tcred->suid) && !uid_eq(cred->euid, tcred->uid) && |
1416 | !uid_eq(cred->uid, tcred->suid) && !uid_eq(cred->uid, tcred->uid) && | |
742755a1 | 1417 | !capable(CAP_SYS_NICE)) { |
c69e8d9c | 1418 | rcu_read_unlock(); |
742755a1 | 1419 | err = -EPERM; |
5e9a0f02 | 1420 | goto out; |
742755a1 | 1421 | } |
c69e8d9c | 1422 | rcu_read_unlock(); |
742755a1 | 1423 | |
86c3a764 DQ |
1424 | err = security_task_movememory(task); |
1425 | if (err) | |
5e9a0f02 | 1426 | goto out; |
86c3a764 | 1427 | |
3268c63e CL |
1428 | task_nodes = cpuset_mems_allowed(task); |
1429 | mm = get_task_mm(task); | |
1430 | put_task_struct(task); | |
1431 | ||
6e8b09ea SL |
1432 | if (!mm) |
1433 | return -EINVAL; | |
1434 | ||
1435 | if (nodes) | |
1436 | err = do_pages_move(mm, task_nodes, nr_pages, pages, | |
1437 | nodes, status, flags); | |
1438 | else | |
1439 | err = do_pages_stat(mm, nr_pages, pages, status); | |
742755a1 | 1440 | |
742755a1 CL |
1441 | mmput(mm); |
1442 | return err; | |
3268c63e CL |
1443 | |
1444 | out: | |
1445 | put_task_struct(task); | |
1446 | return err; | |
742755a1 | 1447 | } |
742755a1 | 1448 | |
7b2259b3 CL |
1449 | /* |
1450 | * Call migration functions in the vma_ops that may prepare | |
1451 | * memory in a vm for migration. migration functions may perform | |
1452 | * the migration for vmas that do not have an underlying page struct. | |
1453 | */ | |
1454 | int migrate_vmas(struct mm_struct *mm, const nodemask_t *to, | |
1455 | const nodemask_t *from, unsigned long flags) | |
1456 | { | |
1457 | struct vm_area_struct *vma; | |
1458 | int err = 0; | |
1459 | ||
1001c9fb | 1460 | for (vma = mm->mmap; vma && !err; vma = vma->vm_next) { |
7b2259b3 CL |
1461 | if (vma->vm_ops && vma->vm_ops->migrate) { |
1462 | err = vma->vm_ops->migrate(vma, to, from, flags); | |
1463 | if (err) | |
1464 | break; | |
1465 | } | |
1466 | } | |
1467 | return err; | |
1468 | } | |
7039e1db PZ |
1469 | |
1470 | #ifdef CONFIG_NUMA_BALANCING | |
1471 | /* | |
1472 | * Returns true if this is a safe migration target node for misplaced NUMA | |
1473 | * pages. Currently it only checks the watermarks which crude | |
1474 | */ | |
1475 | static bool migrate_balanced_pgdat(struct pglist_data *pgdat, | |
3abef4e6 | 1476 | unsigned long nr_migrate_pages) |
7039e1db PZ |
1477 | { |
1478 | int z; | |
1479 | for (z = pgdat->nr_zones - 1; z >= 0; z--) { | |
1480 | struct zone *zone = pgdat->node_zones + z; | |
1481 | ||
1482 | if (!populated_zone(zone)) | |
1483 | continue; | |
1484 | ||
6e543d57 | 1485 | if (!zone_reclaimable(zone)) |
7039e1db PZ |
1486 | continue; |
1487 | ||
1488 | /* Avoid waking kswapd by allocating pages_to_migrate pages. */ | |
1489 | if (!zone_watermark_ok(zone, 0, | |
1490 | high_wmark_pages(zone) + | |
1491 | nr_migrate_pages, | |
1492 | 0, 0)) | |
1493 | continue; | |
1494 | return true; | |
1495 | } | |
1496 | return false; | |
1497 | } | |
1498 | ||
1499 | static struct page *alloc_misplaced_dst_page(struct page *page, | |
1500 | unsigned long data, | |
1501 | int **result) | |
1502 | { | |
1503 | int nid = (int) data; | |
1504 | struct page *newpage; | |
1505 | ||
1506 | newpage = alloc_pages_exact_node(nid, | |
1507 | (GFP_HIGHUSER_MOVABLE | GFP_THISNODE | | |
1508 | __GFP_NOMEMALLOC | __GFP_NORETRY | | |
1509 | __GFP_NOWARN) & | |
1510 | ~GFP_IOFS, 0); | |
bac0382c | 1511 | if (newpage) |
90572890 | 1512 | page_cpupid_xchg_last(newpage, page_cpupid_last(page)); |
bac0382c | 1513 | |
7039e1db PZ |
1514 | return newpage; |
1515 | } | |
1516 | ||
a8f60772 MG |
1517 | /* |
1518 | * page migration rate limiting control. | |
1519 | * Do not migrate more than @pages_to_migrate in a @migrate_interval_millisecs | |
1520 | * window of time. Default here says do not migrate more than 1280M per second. | |
e14808b4 MG |
1521 | * If a node is rate-limited then PTE NUMA updates are also rate-limited. However |
1522 | * as it is faults that reset the window, pte updates will happen unconditionally | |
1523 | * if there has not been a fault since @pteupdate_interval_millisecs after the | |
1524 | * throttle window closed. | |
a8f60772 MG |
1525 | */ |
1526 | static unsigned int migrate_interval_millisecs __read_mostly = 100; | |
e14808b4 | 1527 | static unsigned int pteupdate_interval_millisecs __read_mostly = 1000; |
a8f60772 MG |
1528 | static unsigned int ratelimit_pages __read_mostly = 128 << (20 - PAGE_SHIFT); |
1529 | ||
e14808b4 MG |
1530 | /* Returns true if NUMA migration is currently rate limited */ |
1531 | bool migrate_ratelimited(int node) | |
1532 | { | |
1533 | pg_data_t *pgdat = NODE_DATA(node); | |
1534 | ||
1535 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window + | |
1536 | msecs_to_jiffies(pteupdate_interval_millisecs))) | |
1537 | return false; | |
1538 | ||
1539 | if (pgdat->numabalancing_migrate_nr_pages < ratelimit_pages) | |
1540 | return false; | |
1541 | ||
1542 | return true; | |
1543 | } | |
1544 | ||
b32967ff | 1545 | /* Returns true if the node is migrate rate-limited after the update */ |
d28d4335 | 1546 | bool numamigrate_update_ratelimit(pg_data_t *pgdat, unsigned long nr_pages) |
7039e1db | 1547 | { |
b32967ff | 1548 | bool rate_limited = false; |
7039e1db | 1549 | |
a8f60772 MG |
1550 | /* |
1551 | * Rate-limit the amount of data that is being migrated to a node. | |
1552 | * Optimal placement is no good if the memory bus is saturated and | |
1553 | * all the time is being spent migrating! | |
1554 | */ | |
1555 | spin_lock(&pgdat->numabalancing_migrate_lock); | |
1556 | if (time_after(jiffies, pgdat->numabalancing_migrate_next_window)) { | |
1557 | pgdat->numabalancing_migrate_nr_pages = 0; | |
1558 | pgdat->numabalancing_migrate_next_window = jiffies + | |
1559 | msecs_to_jiffies(migrate_interval_millisecs); | |
1560 | } | |
b32967ff MG |
1561 | if (pgdat->numabalancing_migrate_nr_pages > ratelimit_pages) |
1562 | rate_limited = true; | |
1563 | else | |
d28d4335 | 1564 | pgdat->numabalancing_migrate_nr_pages += nr_pages; |
a8f60772 | 1565 | spin_unlock(&pgdat->numabalancing_migrate_lock); |
b32967ff MG |
1566 | |
1567 | return rate_limited; | |
1568 | } | |
1569 | ||
1570 | int numamigrate_isolate_page(pg_data_t *pgdat, struct page *page) | |
1571 | { | |
340ef390 | 1572 | int page_lru; |
a8f60772 | 1573 | |
3abef4e6 MG |
1574 | VM_BUG_ON(compound_order(page) && !PageTransHuge(page)); |
1575 | ||
7039e1db | 1576 | /* Avoid migrating to a node that is nearly full */ |
340ef390 HD |
1577 | if (!migrate_balanced_pgdat(pgdat, 1UL << compound_order(page))) |
1578 | return 0; | |
7039e1db | 1579 | |
340ef390 HD |
1580 | if (isolate_lru_page(page)) |
1581 | return 0; | |
7039e1db | 1582 | |
340ef390 HD |
1583 | /* |
1584 | * migrate_misplaced_transhuge_page() skips page migration's usual | |
1585 | * check on page_count(), so we must do it here, now that the page | |
1586 | * has been isolated: a GUP pin, or any other pin, prevents migration. | |
1587 | * The expected page count is 3: 1 for page's mapcount and 1 for the | |
1588 | * caller's pin and 1 for the reference taken by isolate_lru_page(). | |
1589 | */ | |
1590 | if (PageTransHuge(page) && page_count(page) != 3) { | |
1591 | putback_lru_page(page); | |
1592 | return 0; | |
7039e1db PZ |
1593 | } |
1594 | ||
340ef390 HD |
1595 | page_lru = page_is_file_cache(page); |
1596 | mod_zone_page_state(page_zone(page), NR_ISOLATED_ANON + page_lru, | |
1597 | hpage_nr_pages(page)); | |
1598 | ||
149c33e1 | 1599 | /* |
340ef390 HD |
1600 | * Isolating the page has taken another reference, so the |
1601 | * caller's reference can be safely dropped without the page | |
1602 | * disappearing underneath us during migration. | |
149c33e1 MG |
1603 | */ |
1604 | put_page(page); | |
340ef390 | 1605 | return 1; |
b32967ff MG |
1606 | } |
1607 | ||
1608 | /* | |
1609 | * Attempt to migrate a misplaced page to the specified destination | |
1610 | * node. Caller is expected to have an elevated reference count on | |
1611 | * the page that will be dropped by this function before returning. | |
1612 | */ | |
1bc115d8 MG |
1613 | int migrate_misplaced_page(struct page *page, struct vm_area_struct *vma, |
1614 | int node) | |
b32967ff MG |
1615 | { |
1616 | pg_data_t *pgdat = NODE_DATA(node); | |
340ef390 | 1617 | int isolated; |
b32967ff MG |
1618 | int nr_remaining; |
1619 | LIST_HEAD(migratepages); | |
1620 | ||
1621 | /* | |
1bc115d8 MG |
1622 | * Don't migrate file pages that are mapped in multiple processes |
1623 | * with execute permissions as they are probably shared libraries. | |
b32967ff | 1624 | */ |
1bc115d8 MG |
1625 | if (page_mapcount(page) != 1 && page_is_file_cache(page) && |
1626 | (vma->vm_flags & VM_EXEC)) | |
b32967ff | 1627 | goto out; |
b32967ff MG |
1628 | |
1629 | /* | |
1630 | * Rate-limit the amount of data that is being migrated to a node. | |
1631 | * Optimal placement is no good if the memory bus is saturated and | |
1632 | * all the time is being spent migrating! | |
1633 | */ | |
340ef390 | 1634 | if (numamigrate_update_ratelimit(pgdat, 1)) |
b32967ff | 1635 | goto out; |
b32967ff MG |
1636 | |
1637 | isolated = numamigrate_isolate_page(pgdat, page); | |
1638 | if (!isolated) | |
1639 | goto out; | |
1640 | ||
1641 | list_add(&page->lru, &migratepages); | |
9c620e2b HD |
1642 | nr_remaining = migrate_pages(&migratepages, alloc_misplaced_dst_page, |
1643 | node, MIGRATE_ASYNC, MR_NUMA_MISPLACED); | |
b32967ff MG |
1644 | if (nr_remaining) { |
1645 | putback_lru_pages(&migratepages); | |
1646 | isolated = 0; | |
1647 | } else | |
1648 | count_vm_numa_event(NUMA_PAGE_MIGRATE); | |
7039e1db | 1649 | BUG_ON(!list_empty(&migratepages)); |
7039e1db | 1650 | return isolated; |
340ef390 HD |
1651 | |
1652 | out: | |
1653 | put_page(page); | |
1654 | return 0; | |
7039e1db | 1655 | } |
220018d3 | 1656 | #endif /* CONFIG_NUMA_BALANCING */ |
b32967ff | 1657 | |
220018d3 | 1658 | #if defined(CONFIG_NUMA_BALANCING) && defined(CONFIG_TRANSPARENT_HUGEPAGE) |
340ef390 HD |
1659 | /* |
1660 | * Migrates a THP to a given target node. page must be locked and is unlocked | |
1661 | * before returning. | |
1662 | */ | |
b32967ff MG |
1663 | int migrate_misplaced_transhuge_page(struct mm_struct *mm, |
1664 | struct vm_area_struct *vma, | |
1665 | pmd_t *pmd, pmd_t entry, | |
1666 | unsigned long address, | |
1667 | struct page *page, int node) | |
1668 | { | |
1669 | unsigned long haddr = address & HPAGE_PMD_MASK; | |
1670 | pg_data_t *pgdat = NODE_DATA(node); | |
1671 | int isolated = 0; | |
1672 | struct page *new_page = NULL; | |
1673 | struct mem_cgroup *memcg = NULL; | |
1674 | int page_lru = page_is_file_cache(page); | |
1675 | ||
b32967ff MG |
1676 | /* |
1677 | * Rate-limit the amount of data that is being migrated to a node. | |
1678 | * Optimal placement is no good if the memory bus is saturated and | |
1679 | * all the time is being spent migrating! | |
1680 | */ | |
d28d4335 | 1681 | if (numamigrate_update_ratelimit(pgdat, HPAGE_PMD_NR)) |
b32967ff MG |
1682 | goto out_dropref; |
1683 | ||
1684 | new_page = alloc_pages_node(node, | |
1685 | (GFP_TRANSHUGE | GFP_THISNODE) & ~__GFP_WAIT, HPAGE_PMD_ORDER); | |
340ef390 HD |
1686 | if (!new_page) |
1687 | goto out_fail; | |
1688 | ||
90572890 | 1689 | page_cpupid_xchg_last(new_page, page_cpupid_last(page)); |
b32967ff MG |
1690 | |
1691 | isolated = numamigrate_isolate_page(pgdat, page); | |
340ef390 | 1692 | if (!isolated) { |
b32967ff | 1693 | put_page(new_page); |
340ef390 | 1694 | goto out_fail; |
b32967ff MG |
1695 | } |
1696 | ||
1697 | /* Prepare a page as a migration target */ | |
1698 | __set_page_locked(new_page); | |
1699 | SetPageSwapBacked(new_page); | |
1700 | ||
1701 | /* anon mapping, we can simply copy page->mapping to the new page: */ | |
1702 | new_page->mapping = page->mapping; | |
1703 | new_page->index = page->index; | |
1704 | migrate_page_copy(new_page, page); | |
1705 | WARN_ON(PageLRU(new_page)); | |
1706 | ||
1707 | /* Recheck the target PMD */ | |
1708 | spin_lock(&mm->page_table_lock); | |
1709 | if (unlikely(!pmd_same(*pmd, entry))) { | |
1710 | spin_unlock(&mm->page_table_lock); | |
1711 | ||
1712 | /* Reverse changes made by migrate_page_copy() */ | |
1713 | if (TestClearPageActive(new_page)) | |
1714 | SetPageActive(page); | |
1715 | if (TestClearPageUnevictable(new_page)) | |
1716 | SetPageUnevictable(page); | |
1717 | mlock_migrate_page(page, new_page); | |
1718 | ||
1719 | unlock_page(new_page); | |
1720 | put_page(new_page); /* Free it */ | |
1721 | ||
a54a407f MG |
1722 | /* Retake the callers reference and putback on LRU */ |
1723 | get_page(page); | |
b32967ff | 1724 | putback_lru_page(page); |
a54a407f MG |
1725 | mod_zone_page_state(page_zone(page), |
1726 | NR_ISOLATED_ANON + page_lru, -HPAGE_PMD_NR); | |
1727 | goto out_fail; | |
b32967ff MG |
1728 | } |
1729 | ||
1730 | /* | |
1731 | * Traditional migration needs to prepare the memcg charge | |
1732 | * transaction early to prevent the old page from being | |
1733 | * uncharged when installing migration entries. Here we can | |
1734 | * save the potential rollback and start the charge transfer | |
1735 | * only when migration is already known to end successfully. | |
1736 | */ | |
1737 | mem_cgroup_prepare_migration(page, new_page, &memcg); | |
1738 | ||
1739 | entry = mk_pmd(new_page, vma->vm_page_prot); | |
1740 | entry = pmd_mknonnuma(entry); | |
1741 | entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma); | |
1742 | entry = pmd_mkhuge(entry); | |
1743 | ||
a54a407f | 1744 | pmdp_clear_flush(vma, haddr, pmd); |
b32967ff | 1745 | set_pmd_at(mm, haddr, pmd, entry); |
a54a407f | 1746 | page_add_new_anon_rmap(new_page, vma, haddr); |
ce4a9cc5 | 1747 | update_mmu_cache_pmd(vma, address, &entry); |
b32967ff MG |
1748 | page_remove_rmap(page); |
1749 | /* | |
1750 | * Finish the charge transaction under the page table lock to | |
1751 | * prevent split_huge_page() from dividing up the charge | |
1752 | * before it's fully transferred to the new page. | |
1753 | */ | |
1754 | mem_cgroup_end_migration(memcg, page, new_page, true); | |
1755 | spin_unlock(&mm->page_table_lock); | |
1756 | ||
1757 | unlock_page(new_page); | |
1758 | unlock_page(page); | |
1759 | put_page(page); /* Drop the rmap reference */ | |
1760 | put_page(page); /* Drop the LRU isolation reference */ | |
1761 | ||
1762 | count_vm_events(PGMIGRATE_SUCCESS, HPAGE_PMD_NR); | |
1763 | count_vm_numa_events(NUMA_PAGE_MIGRATE, HPAGE_PMD_NR); | |
1764 | ||
b32967ff MG |
1765 | mod_zone_page_state(page_zone(page), |
1766 | NR_ISOLATED_ANON + page_lru, | |
1767 | -HPAGE_PMD_NR); | |
1768 | return isolated; | |
1769 | ||
340ef390 HD |
1770 | out_fail: |
1771 | count_vm_events(PGMIGRATE_FAIL, HPAGE_PMD_NR); | |
b32967ff | 1772 | out_dropref: |
a54a407f MG |
1773 | entry = pmd_mknonnuma(entry); |
1774 | set_pmd_at(mm, haddr, pmd, entry); | |
1775 | update_mmu_cache_pmd(vma, address, &entry); | |
1776 | ||
340ef390 | 1777 | unlock_page(page); |
b32967ff | 1778 | put_page(page); |
b32967ff MG |
1779 | return 0; |
1780 | } | |
7039e1db PZ |
1781 | #endif /* CONFIG_NUMA_BALANCING */ |
1782 | ||
1783 | #endif /* CONFIG_NUMA */ |